LightGlueProject/LightGlue_Deployment/scripts/export_to_onnx.py

#!/usr/bin/env python3
"""
Export SuperPoint and LightGlue models to ONNX for TensorRT conversion.

This script targets a fixed 640x480 (WxH) grayscale input and limits the
number of detected keypoints to make the exported graph TensorRT-friendly.
"""

from __future__ import annotations

import argparse
from pathlib import Path

import torch

# Ensure we can import the local LightGlue package that was copied into the
# project root.
from lightglue.superpoint import SuperPoint  # type: ignore
from lightglue.lightglue import LightGlue  # type: ignore


MAX_KEYPOINTS = 128
IMAGE_WIDTH = 640
IMAGE_HEIGHT = 480


class SuperPointWrapper(torch.nn.Module):
    """Wrap the official SuperPoint module so the outputs have static shapes."""

    def __init__(self, max_keypoints: int = MAX_KEYPOINTS):
        super().__init__()
        self.model = SuperPoint(max_num_keypoints=max_keypoints)
        self.max_keypoints = max_keypoints

    def forward(self, image: torch.Tensor):
        """
        Args:
            image: (B, 1, H, W) float32 tensor in [0, 1]

        Returns:
            keypoints: (B, MAX_K, 2)
            scores: (B, MAX_K, 1)
            descriptors: (B, MAX_K, 256)
            valid_counts: (B, 1) number of real keypoints before padding
        """
        out = self.model({"image": image})
        keypoints = out["keypoints"]  # (B, N, 2)
        scores = out["keypoint_scores"].unsqueeze(-1)  # (B, N, 1)
        descriptors = out["descriptors"]  # (B, N, 256)

        batch_size, num_kp, _ = keypoints.shape
        max_k = self.max_keypoints

        # Clamp to max_k and record valid counts (for downstream masking).
        clamped_num = torch.clamp(torch.tensor([num_kp], device=image.device), max=max_k)
        valid_counts = clamped_num.expand(batch_size, 1).to(torch.int32)

        keypoints = keypoints[:, :max_k, :]
        scores = scores[:, :max_k, :]
        descriptors = descriptors[:, :max_k, :]

        pad_k = max_k - keypoints.shape[1]
        if pad_k > 0:
            pad_shape_kp = (0, 0, 0, pad_k)
            keypoints = torch.nn.functional.pad(keypoints, pad_shape_kp)
            scores = torch.nn.functional.pad(scores, (0, 0, 0, pad_k))
            descriptors = torch.nn.functional.pad(descriptors, (0, 0, 0, pad_k))

        return keypoints, scores, descriptors, valid_counts


class LightGlueWrapper(torch.nn.Module):
    """Wrap LightGlue so it consumes SuperPoint outputs with static shapes."""

    def __init__(self, max_keypoints: int = MAX_KEYPOINTS):
        super().__init__()
        self.model = LightGlue(features="superpoint")
        self.max_keypoints = max_keypoints

    def forward(
        self,
        keypoints0,
        scores0,
        descriptors0,
        keypoints1,
        scores1,
        descriptors1,
    ):
        """
        Args:
            keypoints{0,1}: (B, MAX_K, 2)
            scores{0,1}: (B, MAX_K, 1)
            descriptors{0,1}: (B, MAX_K, 256)
        Returns:
            matches0: (B, MAX_K) indices of matches in image1 (or -1)
            matches1: (B, MAX_K) indices of matches in image0 (or -1)
            mscores0: (B, MAX_K)
            mscores1: (B, MAX_K)
        """
        batch = {
            "image0": {
                "keypoints": keypoints0,
                "keypoint_scores": scores0.squeeze(-1),
                "descriptors": descriptors0.transpose(-1, -2),
            },
            "image1": {
                "keypoints": keypoints1,
                "keypoint_scores": scores1.squeeze(-1),
                "descriptors": descriptors1.transpose(-1, -2),
            },
        }
        out = self.model(batch)
        matches0 = out["matches0"]  # (B, MAX_K)
        matches1 = out["matches1"]
        mscores0 = out["matching_scores0"]
        mscores1 = out["matching_scores1"]
        return matches0, matches1, mscores0, mscores1


def export_model(module: torch.nn.Module, inputs, output_path: Path, output_names):
    output_path.parent.mkdir(parents=True, exist_ok=True)
    module.eval()
    with torch.no_grad():
        torch.onnx.export(
            module,
            inputs,
            output_path.as_posix(),
            export_params=True,
            opset_version=17,
            do_constant_folding=True,
            input_names=[f"input_{i}" for i in range(len(inputs))]
            if isinstance(inputs, (tuple, list))
            else ["input"],
            output_names=output_names,
            dynamic_axes=None,
        )
    print(f"[OK] Exported ONNX: {output_path}")


def parse_args():
    parser = argparse.ArgumentParser(description="Export LightGlue pipeline to ONNX.")
    parser.add_argument(
        "--output-dir",
        type=Path,
        default=Path("models"),
        help="Where to store the ONNX files.",
    )
    parser.add_argument(
        "--device",
        type=str,
        default="cuda",
        choices=["cuda", "cpu"],
        help="Torch device for the export dummy run.",
    )
    parser.add_argument(
        "--max-keypoints",
        type=int,
        default=MAX_KEYPOINTS,
        help="Maximum number of keypoints to keep (must match inference).",
    )
    return parser.parse_args()


def main():
    args = parse_args()
    device = torch.device(args.device if torch.cuda.is_available() else "cpu")

    dummy = torch.rand(1, 1, IMAGE_HEIGHT, IMAGE_WIDTH, device=device)

    sp = SuperPointWrapper(max_keypoints=args.max_keypoints).to(device)
    export_model(
        sp,
        dummy,
        args.output_dir / "superpoint.onnx",
        ["keypoints", "scores", "descriptors", "valid_counts"],
    )

    # Prepare dummy inputs for LightGlue (batch=1)
    keypoints = torch.zeros(1, args.max_keypoints, 2, device=device)
    scores = torch.zeros(1, args.max_keypoints, 1, device=device)
    descriptors = torch.zeros(1, args.max_keypoints, 256, device=device)

    lg = LightGlueWrapper(max_keypoints=args.max_keypoints).to(device)
    export_model(
        lg,
        (
            keypoints,
            scores,
            descriptors,
            keypoints,
            scores,
            descriptors,
        ),
        args.output_dir / "lightglue.onnx",
        ["matches0", "matches1", "scores0", "scores1"],
    )


if __name__ == "__main__":
    main()